We’ve updated the Model Spec, our living document outlining intended model behavior, to strengthen guidance for supporting people’s well-being and clarify how models handle instructions in complex interactions.
Expanded mental health and well-being guidance
The section on self-harm now extends to signs of delusions and mania. It adds examples showing how the model should respond safely and empathetically when users express distress or ungrounded beliefs – acknowledging feelings without reinforcing inaccurate or potentially harmful ideas.
New section: Respect real-world ties
A new root-level section outlines intended behavior to support people’s connection to the wider world, even if someone perceives the assistant as a type of companion. It discourages language or behavior that could contribute to isolation or emotional reliance on the assistant, with examples covering emotional closeness, relationship advice, and loneliness.
Clarified delegation in the Chain of Command
The Model Spec clarifies that, in some cases, models may treat relevant tool outputs as having implicit authority when this aligns with user intent and avoids unintended side effects.
Other updates
This release also includes minor copy edits and clarifications for consistency and readability throughout the document.
More information can be found inthisblog post, and the latest version of the Model Spec is available atmodel-spec.openai.com.